Deep learning-based electroencephalogram regulation monitoring data feature mining method and computer system

Abstract

The application provides a deep learning-based electroencephalogram regulation monitoring data feature mining method and a computer system, relates to the technical field of data processing, prunes network configuration variables of an initially adjusted monitoring data feature mining network, and secondarily adjusts the pruned monitoring data feature mining network to obtain a monitoring data feature mining network in a convergent state. The application reduces the size of the obtained monitoring data feature mining network, so as to prevent the monitoring data feature mining network from being difficult to run on a deployed electroencephalogram regulation device.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and computer system for feature mining of electroencephalogram (EEG) modulation monitoring data based on deep learning. Background Technology

[0002] Electroencephalography (EEG) signals contain a wealth of information about brain activity and have enormous application potential in numerous fields, such as medicine, neuroscience research, and brain-computer interface (BCI) technology. In medicine, EEG signals can be used for disease diagnosis, treatment monitoring, and assessment of brain function. For example, neurological diseases such as epilepsy and Alzheimer's disease are often accompanied by characteristic changes in EEG signals; analyzing these signals can assist doctors in early diagnosis and disease monitoring. In neuroscience research, EEG signals contribute to a deeper understanding of complex mechanisms such as cognitive processes and neural plasticity. In BCI technology, EEG signals are a key medium for communication between the brain and external devices, enabling paralyzed patients to control assistive devices. With the application of deep learning technology in processing EEG modulation and monitoring data, the constructed neural network models often require a large scale, containing numerous neurons and connection weights, when handling complex EEG data feature mining tasks. However, in practical applications, especially in portable or resource-constrained devices (such as wearable EEG monitoring devices), limited computing resources (such as memory and processing power) become a significant challenge. Large-scale neural networks can lead to slow operation, excessive power consumption, or even failure to function properly on these devices. Summary of the Invention

[0003] In view of this, the present invention provides a method and computer system for feature mining of EEG modulation monitoring data based on deep learning. The technical solution of the present invention is implemented as follows:

[0004] In a first aspect, embodiments of the present invention provide a method for feature mining of EEG modulation monitoring data based on deep learning. The method includes: acquiring a knowledge template library for the monitoring data feature mining network, the knowledge template library including one or more target-oriented knowledge templates and one or more targetless knowledge templates, wherein the target-oriented knowledge templates represent EEG modulation monitoring data with target features, and the targetless knowledge templates represent EEG modulation monitoring data without the target features; adjusting the monitoring data feature mining network using the knowledge template library to obtain an initially adjusted monitoring data feature mining network; and adjusting the network of the initially adjusted monitoring data feature mining network. Pruning is performed on the configuration variables to obtain a pruned monitoring data feature mining network; wherein the memory usage of the pruned monitoring data feature mining network is less than that of the initially calibrated monitoring data feature mining network; the pruned monitoring data feature mining network is calibrated using the knowledge template library to obtain a converged monitoring data feature mining network; based on the representation information extraction component of the converged monitoring data feature mining network, the data feature vector of the EEG modulation monitoring data to be mined is obtained; based on the data feature vector of the EEG modulation monitoring data to be mined, it is determined whether the target feature is present in the EEG modulation monitoring data to be mined.

[0005] In a second aspect, the present invention provides a computer system including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the method described above.

[0006] The beneficial effects of this invention are as follows: The deep learning-based EEG modulation monitoring data feature mining method and computer system provided by this invention prunes the network configuration variables of the initially calibrated monitoring data feature mining network, and performs secondary calibration on the pruned monitoring data feature mining network to obtain a converged monitoring data feature mining network. This invention reduces the size of the obtained monitoring data feature mining network, thus preventing difficulties in running the monitoring data feature mining network on deployed EEG modulation devices. Attached Figure Description

[0007] Figure 1 This is a schematic diagram illustrating the implementation process of a deep learning-based method for mining features from EEG modulation monitoring data, as provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] This invention provides a deep learning-based method for feature mining of EEG modulation monitoring data, which can be executed by a computer system's processor. The computer system can refer to devices with data processing capabilities, such as servers, laptops, tablets, desktop computers, and EEG modulation devices.

[0011] Figure 1 This is a schematic diagram illustrating the implementation process of a deep learning-based method for feature mining of EEG modulation monitoring data, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0012] Step S100: Obtain the knowledge template library of the monitoring data feature mining network. The knowledge template library includes one or more target knowledge templates and one or more targetless knowledge templates. Target knowledge templates represent EEG modulation monitoring data with target features, and targetless knowledge templates represent EEG modulation monitoring data without target features.

[0013] The knowledge template library contains one or more targeted knowledge templates and one or more non-targeted knowledge templates. Targeted knowledge templates represent EEG modulation monitoring data with the target feature, while non-targeted knowledge templates represent EEG modulation monitoring data without the target feature. To obtain this knowledge template library, an original template library needs to be constructed. Taking the detection of epileptic seizure characteristics in EEG signals as an example, the construction of the original template library comes from a large amount of EEG monitoring data. The original targeted knowledge templates may be extracted from EEG data collected during seizures of patients diagnosed with epilepsy. These data have some obvious characteristics, such as a sudden drop in EEG signal energy in a specific frequency band (e.g., the 8-12 Hz alpha band) and a spike-like energy burst in a high-frequency band (e.g., the 30-80 Hz gamma band). These features can be represented by feature vectors. Let's assume a feature vector is [α-band energy reduction amplitude, γ-band peak count, peak amplitude], where each element represents a specific numerical range, such as [0.3~0.5 (representing the proportion of energy reduction amplitude), 3~5 (number of peaks), 0.8~1.2 (peak amplitude, in microvolts)]. The original targetless knowledge template is obtained from EEG data of healthy individuals or patients without seizures. Their EEG signal feature vector might be [α-band stable energy, low γ-band activity], with corresponding numerical ranges of [0.9~1.1 (α-band energy proportion), 0~1 (quantization value of low-energy γ-band activity)].

[0014] After constructing the initial template library, the computer system needs to process the target features to enrich the library. Assume the target feature set includes features such as frequency stability, phase synchronization, and signal complexity of EEG signals. The computer system will perform combination operations on these target features. For example, frequency stability and signal complexity can be combined into a sequence of data items. For frequency stability, an array can be used to represent frequency fluctuations over different time periods, such as [0.1, 0.05, 0.08] (representing the amplitude of frequency fluctuations over three consecutive time periods). Signal complexity can be represented by a specific numerical value, such as 0.6 (derived based on a certain complexity calculation algorithm). This combined sequence of data items contains the relevant data for these two target features.

[0015] Next, we further manipulate this data sequence. Let's assume we extract a portion of this data sequence as a subsequence, for example, extracting only the signal complexity subsequence. Then, we fuse this subsequence with the original targetless knowledge template. The EEG data in the original targetless knowledge template did not originally contain information related to this specific signal complexity; after fusion, the targetless knowledge template is enhanced.

[0016] If the feature represented by this subsequence belongs to the target feature set, such as signal complexity, then this enhanced template can be considered an enhanced template with targeted knowledge. This means that the template now contains information related to the target feature and can be used as a template with targeted knowledge in subsequent operations.

[0017] Conversely, if the feature represented by the subsequence does not belong to the target feature in the target feature set, then the enhanced template is still regarded as an enhanced targetless knowledge template.

[0018] Finally, the computer system constructs a knowledge template library for the monitoring data feature mining network based on the original template library and the newly generated enhanced targeted knowledge templates and enhanced non-targeted knowledge templates. The templates in this knowledge template library will provide important guidance for subsequent network tuning and other operations. For example, in deep learning-based neural network models, the data features in these templates can be used as training data or reference data. During the training process of the neural network, data from the targeted knowledge templates can serve as positive examples, and data from the non-targeted knowledge templates can serve as negative examples, helping the neural network learn how to distinguish between EEG modulation monitoring data with and without targeted features, thereby accurately mining the targeted features from the data.

[0019] Step S200: Adjust the monitoring data feature mining network using the knowledge template library to obtain the initially adjusted monitoring data feature mining network.

[0020] In step S200, the knowledge template library obtained in step S100 is used, which contains targeted knowledge templates and non-targeted knowledge templates, to play a key role in adjusting the monitoring data feature mining network.

[0021] Suppose the monitoring data feature mining network is a deep neural network model whose purpose is to extract specific features from EEG modulation monitoring data. For example, in EEG signal research, if the goal is to detect features associated with a certain cognitive state (such as a state of focus), the network needs to learn to distinguish the association between different EEG signal patterns and states of focus.

[0022] The targeted knowledge templates in the knowledge template base represent EEG modulation monitoring data with targeted features. Taking the state of focus as an example, the data features in the targeted knowledge templates may manifest as energy changes in EEG signals in specific frequency bands and signal synchronicity between different brain regions. For example, in the frontal lobe, the energy of the alpha band (8~12Hz) EEG signal may be relatively reduced in the state of focus, while the signal synchronicity between the frontal and parietal lobes in the beta band (13~30Hz) is enhanced. These features can be represented by feature vectors, such as [relative value of frontal alpha band energy, frontal-parietal beta band signal synchronicity index], with specific values ​​such as [0.8 (representing the proportion of energy after relative reduction), 0.6 (representing the quantification value of signal synchronicity)].

[0023] Targetless knowledge templates represent EEG modulation monitoring data that lacks targeted features. For focused state detection, data in targetless knowledge templates may originate from EEG monitoring in non-focused states, such as EEG signals in relaxed or distracted states. These signals differ from those in focused states in terms of frequency band energy distribution and brain region synchronicity. For example, in a relaxed state, the energy of the frontal lobe alpha band may be relatively high, while the synchronicity of the beta band signals between brain regions is weak; the corresponding feature vector might be [1.2 (relative value of frontal lobe alpha band energy), 0.3 (synchronicity index of frontal lobe to parietal lobe beta band signals)]. The computer system inputs this data from both targeted and targetless knowledge templates into a monitoring data feature mining network. In a deep neural network model, each layer of the network contains multiple neurons connected by weights. When data from a targeted knowledge template is input, the network processes the data according to a pre-defined activation function (such as the ReLU function). For example, given the feature vector [0.8, 0.6] from the input target knowledge template, the neuron performs a dot product operation between this vector and its own weight vector, then passes it through the ReLU function to obtain the output value, which is then passed to the next layer of neurons. The same operation is performed on data without a target knowledge template.

[0024] By continuously inputting data from the knowledge template base into the network, the weights in the network are adjusted based on the error between the input data and the expected output (the output corresponding to the target knowledge template has the target feature, and the output corresponding to the non-target knowledge template does not have the target feature). This adjustment process is based on the backpropagation algorithm. For example, if the network outputs an incorrect result for data in the target knowledge template, it indicates that the network weights need to be adjusted. The backpropagation algorithm calculates the contribution of each weight to the error and then adjusts the weights based on this contribution. After a large amount of data input and weight adjustment, the network gradually learns the data features in the target and non-target knowledge templates, thereby achieving the calibration of the monitoring data feature mining network. Through this calibration process, the computer system obtains the initially calibrated monitoring data feature mining network. This initially calibrated network has a certain ability to learn and distinguish between the features of target and non-targeted EEG modulation monitoring data, laying the foundation for further optimization (such as network configuration variable pruning in step S300). Through this calibration method, the monitoring data feature mining network can better adapt to the feature mining task of EEG modulation monitoring data, improving the accuracy of the mining results.

[0025] Step S300: Prune the network configuration variables of the initially calibrated monitoring data feature mining network to obtain the pruned monitoring data feature mining network; wherein, the memory usage of the pruned monitoring data feature mining network is less than that of the initially calibrated monitoring data feature mining network.

[0026] In step S300, the network configuration variables of the initially calibrated monitoring data feature mining network are pruned to obtain a pruned monitoring data feature mining network, and the memory usage of the pruned network is less than that of the initially calibrated network.

[0027] In this embodiment of the invention, the monitoring data feature mining network is a deep neural network structure, such as a convolutional neural network (CNN) or a recurrent neural network (RNN). This network consists of multiple network components, each containing numerous parameters, which are the network configuration variables.

[0028] Suppose that a network component in the monitoring data feature mining network is a fully connected layer. The weight tensor in the fully connected layer can be represented as a two-dimensional array, for example, an i*j matrix where i = 5 and j = 4. Each element in the matrix is ​​a weight parameter, such as a value of 0.1, -0.2, etc. This weight tensor determines how the input data is transformed in this layer.

[0029] The computer system operates on this weight tensor. First, it performs eigenvalue decomposition on the weight tensor of the first network component, resulting in the product of a first rotation tensor, an eigenvalue tensor, and a second rotation tensor. This process is similar to eigenvalue decomposition of a matrix. For example, the first rotation tensor is an i*i (here i = 5) matrix, the eigenvalue tensor is an i*j (i = 5, j = 4) matrix, and the second rotation tensor is a j*j (j = 4) matrix.

[0030] To understand this process, let's take a simple 2x2 matrix example and assume a simple weight tensor. After eigenvalue splitting, we get Where Q is the first rotation tensor, R is the eigenvalue tensor, and R is the second rotation tensor.

[0031] This eigenvalue tensor contains multiple eigenvalues. The computer system then determines the target eigenvalues ​​that meet the first requirement. For example, the eigenvalues ​​in the eigenvalue tensor might be: (Here, we assume there are i*j = 4 eigenvalues). Following a certain rule, such as selecting the largest eigenvalues ​​in order of magnitude, let's assume we select the first p = 2 eigenvalues. ) as the target feature value.

[0032] Then, configuration variables corresponding to the target eigenvalues ​​are extracted from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor, respectively, to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor.

[0033] Continuing with the example above, if the first rotation tensor Since there are two target feature values, columns corresponding to these two target feature values ​​are extracted from Q, let's say columns 1 and 3, to obtain the first rotation tensor after pruning. The first rotation tensor after pruning contains j*p (here j = 4, p = 2) weight parameters.

[0034] For the eigenvalue tensor, the original is... After pruning, the following was obtained It contains p*p (p = 2) weight parameters.

[0035] For the second rotation tensor Extract the rows corresponding to the target feature values, assuming they are rows 1 and 2, to obtain the pruned second rotation tensor. = It contains p*j (p = 2, j = 4) weight parameters.

[0036] By pruning the various tensors, the computer system successfully streamlined the configuration variables of the first network component. Similar operations were performed on other network components in the monitoring data feature mining network, resulting in the pruned monitoring data feature mining network. This pruning operation effectively reduces redundant parameters in the network, lowers network complexity, and consequently reduces network memory usage. It also improves network efficiency to some extent without significantly sacrificing the network's ability to mine features from EEG modulation monitoring data, providing a more optimized network structure for subsequent adjustments and other operations.

[0037] Step S400: Adjust the feature mining network of the pruned monitoring data using the knowledge template library to obtain the feature mining network of the monitoring data in a converged state.

[0038] In step S400, the computer system adjusts the pruned monitoring data feature mining network through a knowledge template library to obtain a converged monitoring data feature mining network.

[0039] In the deep learning-based EEG modulation monitoring data feature mining task, the monitoring data feature mining network has already undergone certain adjustments in structure and parameters after the previous steps (such as initial tuning and pruning). Although the pruned monitoring data feature mining network reduces memory usage, it may not yet have reached its optimal state and still needs further tuning to achieve convergence.

[0040] Suppose that the monitoring data feature mining network is a multilayer perceptron (MLP) neural network used to extract specific physiological or pathological features from EEG modulation monitoring data. For example, the goal is to extract features related to epileptic seizures from EEG signals, such as energy changes in specific frequency bands of EEG signals and signal complexity.

[0041] The knowledge template base contains both targeted and non-targeted knowledge templates. Taking epilepsy detection as an example, the data in the targeted knowledge templates may come from EEG monitoring data during seizures of patients diagnosed with epilepsy. This data has specific feature vectors, which can be represented as an array, such as [the increase in EEG signal energy in a specific frequency band (e.g., 3-8Hz) is 0.5, and the signal complexity reaches 0.8 (based on a certain complexity calculation method)]. The data in the non-targeted knowledge templates may come from EEG monitoring data of healthy individuals, and its feature vector may be [the corresponding frequency band EEG signal energy is stable, and the signal complexity is 0.2].

[0042] The computer system inputs data from a knowledge template base into a feature mining network for pruned monitoring data. For a multilayer perceptron network, the input layer receives feature vector data from the knowledge template base, which is then processed through hidden layers. Neurons in the hidden layers perform weighted summation of the input data and apply activation functions (such as the sigmoid function or the ReLU function) for nonlinear transformation.

[0043] For example, when the input contains data [0.5, 0.8] from the target knowledge template, assuming the first hidden layer has 3 neurons, each neuron has connection weights with the input layer. For the first neuron, the connection weight vector is assumed to be [0.1, 0.2]. The weighted sum is then 0.1 × 0.5 + 0.2 × 0.8 = 0.21, and then passed through the sigmoid function. The output value is obtained, where sigmoid(0.21)≈0.55. This output value will continue to be passed to the next layer of neurons or used directly as the output (if it is the last hidden layer).

[0044] During backpropagation, the network adjusts its weights based on the error between the output and the expected result (a target knowledge template corresponds to data with the target feature, while a lack of a target knowledge template corresponds to data without the target feature). If the output indicates that the network incorrectly identifies data with the target knowledge template (i.e., classifies it as lacking the target feature), then the weights need to be adjusted to reduce this error. Assuming the error between the output and the expected result after a single input of data with the target knowledge template is E, the contribution of each weight to the error is calculated. For example, for the connection weights 0.1 and 0.2 of the first neuron mentioned earlier, their adjustment amounts are calculated using a specific backpropagation algorithm (such as gradient descent). and Then update the weights to... , .

[0045] The computer system continuously inputs data from the knowledge template base into the pruned monitoring data feature mining network, performing forward and backward propagation operations. As the number of iterations increases, the network weights are continuously adjusted. During this process, the error between the network's output and the expected result gradually decreases. When the error decreases to a certain extent, for example, after multiple iterations, the change in error is less than a preset minimum value (e.g., ...). When the convergence of the monitoring data feature mining network is achieved, it can be considered that the network has reached a convergent state. At this point, the computer system obtains the converged monitoring data feature mining network. This converged network has a good ability to distinguish between target and non-target features in the EEG modulation monitoring data, and can more accurately mine features from the EEG modulation monitoring data, laying a solid foundation for subsequent feature mining of EEG modulation monitoring data based on this network.

[0046] Step S500: The representation information extraction component of the monitoring data feature mining network based on the convergence state obtains the data feature vector of the EEG modulation monitoring data to be mined, and determines whether the EEG modulation monitoring data to be mined has the target feature based on the data feature vector of the EEG modulation monitoring data to be mined.

[0047] In step S500, the computer system obtains the data feature vector of the EEG modulation monitoring data to be mined by the representation information extraction component of the monitoring data feature mining network based on the convergence state, and then determines whether the EEG modulation monitoring data to be mined has the target feature based on the data feature vector.

[0048] In the context of deep learning-based EEG modulation monitoring data feature mining, the converged monitoring data feature mining network has undergone a series of calibration processes (such as initial calibration through a knowledge template base and subsequent calibration after pruning), and is capable of extracting effective features from EEG modulation monitoring data.

[0049] Assuming the convergent monitoring data feature mining network is a deep neural network (e.g., a convolutional neural network), its representation information extraction component can be viewed as a part of the network specifically dedicated to extracting data feature representations. This part may consist of the last few layers of the network, which transform the input EEG modulation monitoring data into a data feature vector.

[0050] For example, EEG modulation monitoring data can be multi-channel EEG signal data collected over a period of time. The data in each channel may contain information such as the intensity of EEG signals in different frequency bands (e.g., alpha, beta, gamma bands). Assuming that the data in each channel is considered as an array, a simple example of three-channel EEG data is as follows: the data array for channel 1 is [alpha band intensity value 0.3, beta band intensity value 0.5, gamma band intensity value 0.2], the data array for channel 2 is [alpha band intensity value 0.4, beta band intensity value 0.3, gamma band intensity value 0.3], and the data array for channel 3 is [alpha band intensity value 0.2, beta band intensity value 0.4, gamma band intensity value 0.4].

[0051] When these EEG modulation monitoring data are input into a converged monitoring data feature mining network, the network's representation information extraction component processes the data. It may extract local features from different channels of data through convolutional layers, then perform feature compression through pooling layers, and finally combine these features into a data feature vector through fully connected layers. Assuming the resulting data feature vector is [0.1, 0.2, 0.3, 0.4], each value represents a quantized value associated with a different feature after network processing.

[0052] The computer system uses this data feature vector to determine whether the target feature exists in the EEG modulation monitoring data to be mined. Assume the target feature is an EEG signal feature related to a certain cognitive state (such as a state of focus). The set of target features may contain multiple sequences of target data items and a sequence of data item attributes corresponding to each sequence of target data items.

[0053] For example, one of the target data item sequences might be a sequence about the energy ratio of EEG signal frequency bands, such as [the energy ratio of the α band should be low in a focused state, and the energy ratio of the β band should be high in a focused state]. The corresponding data item attribute sequence might be [the threshold range of the energy ratio, the threshold range of the energy ratio], assuming it is [α band: 0~0.3, β band: 0.6~1].

[0054] The computer system matches the data feature vector [0.1, 0.2, 0.3, 0.4] with the target data item sequence in the target feature set. If the values ​​in the data feature vector match the conditions in the target data item sequence—for example, assuming the first value in the data feature vector, 0.1, represents the α-band energy ratio and 0.2 represents the β-band energy ratio, and they fall within the threshold range of the attribute sequence corresponding to the target data item sequence—then it can be determined that the EEG modulation monitoring data to be mined contains the target feature (focus-related feature).

[0055] For example, if the target data item sequence is about the synchronicity between different brain regions of EEG signals, such as [the frontal and parietal lobes have high β-band synchronicity under focused conditions], the corresponding attribute sequence is [the threshold range of the synchronicity index, such as 0.6~1]. If a certain value in the data feature vector indicates that the β-band synchronicity index of the frontal and parietal lobes is 0.8, which falls within the threshold range, this also indicates that the target feature is to be mined from the EEG modulation monitoring data.

[0056] In this way, the computer system compares and matches the data feature vectors extracted by the convergent monitoring data feature mining network with the target feature set, thereby accurately determining whether the target feature exists in the EEG modulation monitoring data to be mined. This is of great significance for tasks such as feature analysis, disease diagnosis, or cognitive state assessment in EEG modulation monitoring.

[0057] As one implementation method, step S300, pruning the network configuration variables of the initially calibrated monitoring data feature mining network to obtain the pruned monitoring data feature mining network, may include:

[0058] Step S310: For the first network component in the monitoring data feature mining network after initial calibration, perform eigenvalue splitting on the weight tensor of the first network component to obtain the product of the first rotation tensor, the eigenvalue tensor, and the second rotation tensor.

[0059] Step S320: Among the eigenvalues ​​of the eigenvalue tensor, determine the target eigenvalue that meets the first requirement;

[0060] Step S330: Extract configuration variables corresponding to the target feature values ​​from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor; wherein, the configuration variables of the first network component in the pruned monitoring data feature mining network are composed of the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor.

[0061] A weight tensor is a multidimensional array used to represent the strength of connections between neurons. In this scenario, for the first network component (e.g., a fully connected layer), the weight tensor is an i*j matrix (where i and j represent the dimensionality-related values ​​of the input and output, respectively). For a simple example, if the input layer has 4 neurons and the output layer has 3 neurons, the weight tensor might be a 4*3 matrix. Each element (the weight parameter) determines the signal strength transmitted from the input neuron to the output neuron. For example, the elements in the weight tensor... This represents the connection weights from the i-th input neuron to the j-th output neuron. The weight tensor plays a crucial role in the forward propagation of the neural network. When input data passes through the first network component, each element of the input data is multiplied by the corresponding weight in the weight tensor, and then summed, etc., to transform the input data into a form that the next layer can process. It determines how the network performs linear transformations on the input data and is an important foundation for the neural network to learn and represent data features. Different weight values ​​lead to different propagation paths and results of the input data in the network. By adjusting the weight parameters in the weight tensor, the network can learn different input-output mapping relationships, thereby enabling the mining of features from EEG modulation monitoring data.

[0062] The First Rotation Tensor is an i*i matrix obtained by eigenvalue decomposition of the weight tensor (where i is related to one dimension of the weight tensor). For example, when the weight tensor is 4*3, eigenvalue decomposition might result in a 4*4 matrix as the First Rotation Tensor. Its elements represent a rotation relationship in the feature space. Within the overall framework of eigenvalue decomposition, the First Rotation Tensor transforms the space of the original weight tensor. It works in conjunction with the eigenvalue tensor and the Second Rotation Tensor to re-represent the structure of the weight tensor. In data processing, it helps decompose the complex linear transformation represented by the original weight tensor into a more easily understood and manipulated form. For example, it can represent the complex relationships between different dimensions of the original weight tensor in a new feature space through rotation operations, providing a new perspective for subsequent determination of important eigenvalues ​​and pruning operations.

[0063] An eigenvalue tensor is an i*j matrix (where i and j are related to the dimensions of the weight tensor). Its diagonal elements are eigenvalues, and off-diagonal elements may be 0 (in ideal eigenvalue decomposition). For example, in the case of a 4*3 weight tensor, the eigenvalue tensor might be a 4*3 matrix, where the diagonal eigenvalues ​​represent the scaling factor along the corresponding eigenvector direction. The eigenvalues ​​represent the importance (in the form of scaling factors) of each eigendirection after the rotation operation of the first rotation tensor. Larger eigenvalues ​​mean that the changes in the data have a greater impact on the linear transformation represented by the weight tensor along the corresponding eigenvector direction. By analyzing the eigenvalues ​​in the eigenvalue tensor, we can determine which eigendirections are more critical to the transformation represented by the weight tensor, thus providing a basis for determining the target eigenvalues ​​for pruning operations.

[0064] The Second Rotation Tensor is a j*j matrix (j is related to one dimension of the weight tensor). For example, when the weight tensor is 4*3, the Second Rotation Tensor might be a 3*3 matrix. Similar to the First Rotation Tensor, it represents a rotation operation in the feature space, but from a different dimension. The Second Rotation Tensor works in conjunction with the First Rotation Tensor and the eigenvalue tensor to provide a complete eigenvalue decomposition representation of the weight tensor. It plays a crucial role in decomposing the weight tensor into a more easily analyzed and manipulated form. Specifically, together with the First Rotation Tensor, it reconstructs the representation of the weight tensor through multiplication with the eigenvalue tensor, allowing the linear transformation represented by the weight tensor to be understood from different perspectives (i.e., by analyzing the eigenvalues ​​and the corresponding rotation relationships), providing a foundation for subsequent eigenvalue-based pruning operations.

[0065] The product of the first rotation tensor, the eigenvalue tensor, and the second rotation tensor refers to the result obtained by multiplying these three tensors according to a specific multiplication rule. This result should be equivalent to the original weight tensor (theoretically, numerical errors in the calculation process are ignored). Mathematically, if the first rotation tensor is Q and the eigenvalue tensor is... If the second rotation tensor is R, then the product is: This result should be equal to the original weight tensor W, i.e., W = The purpose of this multiplication result is to verify the correctness of the eigenvalue splitting operation. If the multiplication result can accurately reconstruct the original weight tensor, it indicates that the eigenvalue splitting operation was successful. Simultaneously, this multiplication result also provides a complete framework for subsequent pruning operations based on eigenvalues ​​and related tensors. By decomposing the weight tensor into a multiplicative form of these three tensors, it is easier to analyze and manipulate the elements in each tensor (especially the parts related to eigenvalues), thereby achieving pruning of network configuration variables, reducing network memory usage, and improving network efficiency.

[0066] The computer system performs eigenvalue splitting on the weight tensor of the first network component in the initially calibrated monitoring data feature mining network, and obtains the product of the first rotation tensor, the eigenvalue tensor, and the second rotation tensor.

[0067] Within the neural network framework of deep learning, the monitoring data feature mining network is a complex structure, where the first network component can be, for example, a fully connected layer. The weight tensors in a fully connected layer are a crucial component, determining the connection strength between neurons and how data is transformed during inter-layer transmission.

[0068] Suppose the weight tensor of this fully connected layer is an i*j matrix (where i = 4, j = 3), and each element in the matrix is ​​a weight parameter. Here's a simple example of a weight tensor. .

[0069] The computer system performs eigenvalue decomposition on the weight tensor W. This eigenvalue decomposition is similar to matrix factorization in linear algebra, aiming to decompose the weight tensor into a more manageable and analyzable form. Using specific algorithms (such as singular value decomposition or eigenvalue decomposition), the weight tensor W is decomposed into a first rotation tensor Q and eigenvalue tensors. And the second rotation tensor R, such that W = .

[0070] Suppose that the first rotation tensor Q is calculated to be an i*i (here i = 4) matrix, for example... ; Eigenvalue tensor Given an i*j (i = 4, j = 3) matrix, for example... The second rotation tensor R is a j*j (j = 3) matrix, for example... This yields the product of the first rotation tensor, the eigenvalue tensor, and the second rotation tensor, which corresponds to the original weight tensor W, providing a foundation for subsequent pruning operations.

[0071] The computer system determines the target eigenvalue that meets the first requirement from the eigenvalues ​​of the eigenvalue tensor. Continuing with the example above, the eigenvalue tensor... The eigenvalues ​​in the matrix are 2.0, 1.5, and 1.0 respectively (the case where the last row is all 0 is ignored here, because in actual eigenvalue decomposition, there may be some eigenvalues ​​of 0 that are relatively unimportant compared to the main eigenvalues).

[0072] Determining the target feature values ​​that meet the first requirement requires a clear rule or standard. This rule can be based on various factors, such as the magnitude of the feature values ​​and their influence on the network output. Let's assume the first requirement here is to select larger feature values, because larger feature values ​​often represent principal component directions that have a greater impact on data transformation.

[0073] In this example, 2.0 and 1.5 are relatively large in order of magnitude, so the computer system identifies 2.0 and 1.5 as the target feature values. The purpose of this step is to filter out the feature values ​​in the feature tensor after weight tensor decomposition that are important to network performance or data representation, so that in subsequent pruning operations, network configuration variables related to these important feature values ​​are retained while those that are relatively unimportant are removed, thereby simplifying the network and reducing memory usage.

[0074] In step S330, the computer system extracts configuration variables corresponding to the target eigenvalues ​​from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor.

[0075] For the first rotation tensor Q, since there are two target eigenvalues ​​(2.0 and 1.5), and Q is a 4x4 matrix, we need to extract configuration variables from Q corresponding to these two target eigenvalues. Assume the correspondence here is based on the order of the eigenvalues ​​in the eigenvalue tensor, corresponding to columns in the first rotation tensor. Then, we extract the first and second columns from the first rotation tensor Q, resulting in the pruned first rotation tensor Q', which is a 4x2 matrix. For example... .

[0076] For eigenvalue tensors The original matrix was a 4* matrix. Because it has two target eigenvalues, the pruned eigenvalue tensor is... Only the parts related to these two target feature values ​​are retained, resulting in a 2x2 matrix, for example. .

[0077] For the second rotation tensor R, which is a 3* matrix, according to the correspondence with the target eigenvalues ​​(assuming the order of the eigenvalues ​​in the eigenvalue tensor corresponds to the rows in the second rotation tensor), the first and second rows are extracted, resulting in the pruned second rotation tensor R', which is a 2* matrix. For example... .

[0078] By extracting configuration variables corresponding to the target feature values ​​from the first rotation tensor, eigenvalue tensor, and second rotation tensor, the computer system obtains the pruned first rotation tensor, pruned eigenvalue tensor, and pruned second rotation tensor. These pruned tensors constitute the configuration variables of the pruned first network component, thus realizing the pruning operation on the first network component. Other network components in the monitoring data feature mining network can also be pruned using a similar method, ultimately resulting in the entire pruned monitoring data feature mining network. This pruning operation effectively reduces redundant information in the network, lowers network complexity, and reduces memory usage, while maintaining, to some extent, the network's ability to mine features from EEG modulation monitoring data, providing a more optimized network structure for subsequent network tuning and data feature mining tasks.

[0079] As one implementation, step S320, determining the target eigenvalue that meets the first requirement among the eigenvalues ​​of the eigenvalue tensor, may include:

[0080] Step S321: Sort the eigenvalues ​​of the eigenvalue tensor in descending order to obtain the eigenvalue vector;

[0081] Step S322: Determine the first p eigenvalues ​​in the eigenvalue vector to obtain p target eigenvalues, where p ≥ 2.

[0082] In step S320, the computer system aims to determine the target eigenvalue that meets the first requirement from the eigenvalues ​​of the eigenvalue tensor. This process is achieved through sub-steps S321 to S322.

[0083] In step S321, the computer system sorts the eigenvalues ​​of the eigenvalue tensor in descending order to obtain the eigenvalue vector.

[0084] In the neural network architecture of deep learning, assuming that in the previous step S310, the eigenvalue tensor was obtained by eigenvalue splitting the weight tensor of the first network component (e.g., a specific hidden layer). Let this eigenvalue tensor be an i*j matrix (for simplicity, let i = 5, j = 3), for example... (The last two lines being all 0 is a possible scenario in eigenvalue decomposition; the actual effective eigenvalues ​​are the first three.)

[0085] The computer system extracts the non-zero eigenvalues ​​from this eigenvalue tensor, namely 3.0, 2.5, and 1.8. These eigenvalues ​​are then sorted in descending order. The sorted result forms an eigenvalue vector. In this example, the eigenvalue vector is [3.0, 2.5, 1.8]. This eigenvalue vector represents the sequence of eigenvalues ​​extracted from the eigenvalue tensor according to their importance (here, importance is represented by numerical value, with larger values ​​indicating greater importance). The significance of this sorting operation is that it provides an ordered basis for subsequently determining the target eigenvalues, enabling the computer system to determine the eigenvalues ​​that have a significant impact on the network structure and performance based on certain criteria (such as selecting the first few larger eigenvalues).

[0086] In step S322, the computer system determines the first p eigenvalues ​​in the eigenvalue vector to obtain p target eigenvalues, where p ≥ 2.

[0087] Continuing with the example above, suppose p = 2 is set based on network structure or experience. Then the computer system will select the first two feature values, 3.0 and 2.5, from the obtained feature vector [3.0, 2.5, 1.8] as the target feature values.

[0088] In deep learning-based EEG modulation monitoring data feature mining networks, the method for determining target feature values ​​is based on the importance distribution of feature values ​​within the network. Larger feature values ​​often correspond to directions that have a greater impact on data transformation. For example, during the forward propagation of a neural network, the weight tensor performs a linear transformation on the input data, and the feature values ​​act as scaling factors in this transformation. Larger feature values ​​mean that changes in the input data will be amplified or reduced to a greater extent in the corresponding feature vector direction, thus having a more significant impact on the network's output.

[0089] To illustrate with a simple neural network example, suppose the input EEG modulation monitoring data, after being processed through a series of hidden layers, reaches this first network component undergoing pruning. If we consider the input data as a vector... The output is obtained after a linear transformation of the weight tensor W (where W is the tensor representation related to the previous eigenvalue decomposition). In this process, the eigenvalues ​​in the eigenvalue tensor affect the result of this linear transformation. For example, when the eigenvalues ​​are large, the input vector... The components in the corresponding feature vector direction will be changed to a greater extent, thus affecting the final output. It has a significant impact.

[0090] Therefore, by determining the first p feature values ​​as target feature values, the computer system can retain network configuration variables related to these important feature values ​​in subsequent pruning operations. This reduces network memory usage while preserving the network's ability to mine features from EEG modulation monitoring data as much as possible. These target feature values ​​will serve as the basis for extracting corresponding configuration variables from the first rotation tensor, feature value tensor, and second rotation tensor in subsequent steps (such as step S330), thereby achieving precise pruning of network configuration variables.

[0091] As one implementation, the weight tensor of the first network component includes i*j weight parameters, the first rotation tensor includes i*i weight parameters, the eigenvalue tensor includes i*j weight parameters, and the second rotation tensor includes j*j weight parameters. The number of target eigenvalues ​​is p, where i≥2; j≥2; p≥2. Based on this, step S330, extracting configuration variables corresponding to the target eigenvalues ​​from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor respectively, to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor, may include:

[0092] Step S331: Based on the target feature value, obtain the p column weight parameters corresponding to the target feature value in the first rotation tensor to obtain the pruned first rotation tensor. The pruned first rotation tensor includes i*p weight parameters.

[0093] Step S332: Based on the target feature value, obtain the pruned feature value tensor, which includes p*p weight parameters;

[0094] Step S333: Based on the target feature value, obtain the p rows of weight parameters corresponding to the target feature value in the second rotation tensor to obtain the pruned second rotation tensor. The pruned second rotation tensor includes p*j weight parameters. The configuration variable in each feature value tensor corresponds to a column of weight parameters in the first rotation tensor and a row of weight parameters in the second rotation tensor.

[0095] In step S330, configuration variables corresponding to the target eigenvalues ​​are extracted from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor. This process is implemented through sub-steps S331 to S333.

[0096] In step S331, the computer system, based on the target feature values, obtains the p columns of weight parameters corresponding to the target feature values ​​in the first rotation tensor, thus obtaining the pruned first rotation tensor. The pruned first rotation tensor includes i*p weight parameters. Assuming that in the previous steps, the first rotation tensor Q was an i*i matrix, here we assume i = 5, then... .

[0097] In step S322, the number of target feature values ​​p is determined, assuming p = 3. And it is assumed that the target feature values ​​are selected from the sorted feature value vector, for example, corresponding to the 1st, 3rd, and 4th feature values ​​in the feature value tensor (this is just an example to illustrate the correspondence).

[0098] Since there is a certain correspondence between the position of each eigenvalue in the eigenvalue tensor and the column in the first rotation tensor (this correspondence is determined by the nature of the eigenvalue splitting algorithm), the computer system extracts columns from the first rotation tensor based on this correspondence. Here, columns 1, 3, and 4 are extracted to obtain the pruned first rotation tensor Q'.

[0099] The first rotation tensor Q' after pruning is an i*p matrix (where i = 5, p = 3), i.e. This matrix contains j*p (here j = 5) weight parameters.

[0100] This pruning operation of the first rotation tensor helps reduce network complexity. For example, in a deep neural network used for classifying EEG modulation monitoring data, the first network component might be part of a fully connected layer, and the elements (weight parameters) of the first rotation tensor determine how the data is transformed within that component. By extracting the columns corresponding to the target feature values, it's equivalent to removing connections that have a relatively small impact on data transformation. This is similar to removing some circuit connections with little impact on the overall circuit function in a complex circuit network, thus simplifying the circuit structure. In neural networks, the simplified structure can reduce computation and memory usage while maintaining, to some extent, the network's ability to mine data features.

[0101] In step S332, the computer system obtains the pruned feature value tensor based on the target feature value. The pruned feature value tensor includes p*p weight parameters.

[0102] Assuming the previous eigenvalue tensor It is an i*j matrix, where i = 5 and j = 3, for example... ,in These are valid eigenvalues.

[0103] Since the number of target feature values ​​is determined to be p = 3, and it is assumed that the target feature values ​​are... The computer system will construct a pruned feature tensor based on the target feature values. .

[0104] Eigenvalue tensors after pruning It is a p*p matrix (here p = 3), that is... This matrix contains p*p weight parameters.

[0105] In neural network operations, the eigenvalues ​​in the eigenvalue tensor reflect the scaling ratio of the data in each feature direction after a linear transformation of the weight tensor. Larger eigenvalues ​​mean that changes in the data in the corresponding feature direction have a greater impact on the final result. By pruning the eigenvalue tensor and retaining the parts relevant to the target eigenvalues, it is possible to reduce the network size while capturing the feature directions that have a significant impact on network performance. For example, in feature mining of EEG modulation monitoring data, if the EEG signal is considered as data with multiple feature dimensions, the eigenvalues ​​in the eigenvalue tensor determine the relative importance of different feature dimensions during network propagation. Pruning the eigenvalue tensor focuses on those important feature dimensions identified as target eigenvalues, helping to improve the efficiency and accuracy of the network in processing EEG modulation monitoring data.

[0106] In step S333, the computer system obtains the p rows of weight parameters corresponding to the target feature value in the second rotation tensor based on the target feature value, and obtains the pruned second rotation tensor. The pruned second rotation tensor includes p*j weight parameters.

[0107] Suppose the second rotation tensor R is a j*j matrix, where j = 3. .

[0108] Similarly, based on the number of target eigenvalues ​​p = 3, and the correspondence between the target eigenvalues ​​and the rows in the second rotation tensor (determined by the eigenvalue splitting algorithm), the computer system extracts rows from the second rotation tensor. Here, rows 1, 2, and 3 are extracted to obtain the pruned second rotation tensor R'.

[0109] The second rotation tensor R' after pruning is a p*j matrix (where p = 3, j = 3), i.e. This matrix contains p*j weight parameters.

[0110] In neural networks, the second rotation tensor, similar to the first rotation tensor, participates in the transformation of input data. Pruning the second rotation tensor, removing rows related to non-target features, simplifies the network structure. For example, in a convolutional neural network extracting features from EEG modulation monitoring data, the network component containing the second rotation tensor might be performing spatial or feature dimension transformations on the data. The pruned second rotation tensor reduces unnecessary computation, allowing the network to focus more quickly on important features when processing EEG modulation monitoring data, improving network efficiency, and preserving the network's ability to extract target features while reducing memory usage.

[0111] Through steps S331-S333, the computer system performs pruning operations on the first rotation tensor, the eigenvalue tensor, and the second rotation tensor, respectively, obtaining pruned tensors. These pruned tensors constitute the configuration variables of the pruned first network component, realizing the pruning of the network configuration variables of the first network component. Other network components in the monitoring data feature mining network can also be pruned using a similar method, ultimately obtaining the entire pruned monitoring data feature mining network. This helps reduce network complexity and memory usage, while maintaining the network's ability to mine features from EEG modulation monitoring data to a certain extent.

[0112] As one implementation method, step S100, obtaining the knowledge template library of the monitoring data feature mining network, may include:

[0113] Step S110: Obtain the original template library of the monitoring data feature mining network. The original template library includes one or more original target knowledge templates and one or more original non-target knowledge templates.

[0114] Step S120: Sample one or more target features from the target feature set to be mined, and combine one or more target features to obtain the first data item sequence;

[0115] Step S130: Sample a subsequence from the first data item sequence and add the subsequence to the original targetless knowledge template to obtain an enhanced template; wherein, the subsequence is one or more uninterrupted data items in the first data item sequence;

[0116] Step S140: If the subsequence is a target feature in the target feature set, use the enhanced template as the enhanced target knowledge template;

[0117] Step S150: If the subsequence is not a target feature in the target feature set, use the enhanced template as the enhanced targetless knowledge template;

[0118] Step S160: Based on the original template library, the enhanced targeted knowledge templates, and the enhanced non-targeted knowledge templates, obtain the knowledge template library for the monitoring data feature mining network.

[0119] In step S110, the computer system acquires the original template library of the monitoring data feature mining network. The original template library includes one or more original target knowledge templates and one or more original non-target knowledge templates.

[0120] In this embodiment of the invention, the construction of the original template library is the foundation for the construction of the entire knowledge template library. For example, in the case of using electroencephalogram (EEG) signal analysis to detect epilepsy, the original targeted knowledge templates can be obtained from the EEG monitoring data of patients already diagnosed with epilepsy. These data have specific characteristic patterns, which may manifest as abnormal EEG signals in specific frequency bands during seizures.

[0121] Taking a simple frequency characteristic as an example, EEG signals are divided into different frequency bands, such as the alpha band (8~12Hz), beta band (13~30Hz), and gamma band (30~80Hz). During an epileptic seizure, there may be a sudden increase in gamma band energy, while the alpha band energy relatively decreases. Then, a feature vector in the original targeted knowledge template can be represented as an array: [alpha band energy relative value 0.3 (assuming normal is 1), gamma band energy relative value 1.5 (assuming normal is 1)], where the values ​​are the energy ratios relative to the normal state. This feature vector represents the EEG signal characteristics related to epileptic seizures and is part of the original targeted knowledge template.

[0122] The initial targetless knowledge templates can be obtained from EEG monitoring data of healthy individuals or patients in non-epileptic states. For example, the EEG signals of healthy individuals are relatively stable across various frequency bands, and their feature vectors might be [α band energy relative value 0.9, γ band energy relative value 1.1]. These feature vectors collectively constitute the initial targetless knowledge templates. The initial template library contains multiple such initial targeted and targetless knowledge templates, which serve as initial representations of the characteristics of different types of EEG modulation monitoring data, providing basic materials for subsequently constructing a richer knowledge template library.

[0123] In step S120, the computer system samples one or more target features from the target feature set to be mined, and combines one or more target features to obtain the first data item sequence.

[0124] Continuing with the example of using EEG signals for epilepsy detection, the target feature set to be mined may include frequency characteristics of the EEG signals, peak and trough morphological characteristics of the signals, and signal synchronicity characteristics between different brain regions. The computer system samples from this target feature set.

[0125] Assume that sampling yields frequency characteristics and signal peak and trough morphological characteristics. For frequency characteristics, consider the energy proportions of different frequency bands mentioned earlier; for signal peak and trough morphological characteristics, these can be described by indicators such as peak height, trough depth, and the distance between peaks and troughs. For example, the relative value of peak height is 1.2 (assuming a standard height of 1), the relative value of trough depth is 0.8, and the relative value of peak and trough distance is 1.1 (these relative values ​​are all relative to a certain standard scale).

[0126] These two target features are then combined to obtain the first data item sequence. This first data item sequence can be represented as an array containing multiple elements, such as [α-band energy relative value 0.3, γ-band energy relative value 1.5, peak height relative value 1.2, trough depth relative value 0.8, peak-trough spacing relative value 1.1]. This sequence combines different target features, providing a richer source of information for the subsequent construction of a knowledge template base. This combination method can integrate information from multiple target features, helping to more comprehensively describe the feature patterns in EEG modulation monitoring data. Especially when processing complex EEG signals, the combination of multiple features can improve the ability to represent target features (such as features related to epileptic seizures).

[0127] In step S130, the computer system samples a subsequence from the first data item sequence and adds the subsequence to the original targetless knowledge template to obtain an enhanced template.

[0128] Taking the previously constructed first data item sequence [α-band energy relative value 0.3, γ-band energy relative value 1.5, peak height relative value 1.2, valley depth relative value 0.8, peak-valley spacing relative value 1.1] as an example, the computer system samples a subsequence from it. Assume the sampled subsequence is [peak height relative value 1.2, valley depth relative value 0.8].

[0129] The original targetless knowledge template is obtained from EEG monitoring data of healthy individuals or patients in non-epileptic states, such as the aforementioned feature vector [relative energy value of α band 0.9, relative energy value of γ band 1.1]. The computer system adds the sampled subsequence [relative peak height value of 1.2, relative valley depth value of 0.8] to this original targetless knowledge template.

[0130] The resulting enhanced template has the following characteristics: [α-band energy relative value 0.9, γ-band energy relative value 1.1, peak height relative value 1.2, trough depth relative value 0.8]. This enhanced template incorporates sub-sequence features sampled from the first data item sequence, enriching the template's content. The purpose of this operation is to modify the feature representation of the original targetless knowledge template by introducing sub-sequences that may be related to the target features. This allows the template library to contain more diverse feature combinations, improving the ability to distinguish between different types of EEG modulation monitoring data. In particular, this enhanced template provides more detailed distinguishing criteria when differentiating between data with target features (such as features related to epileptic seizures) and data without target features.

[0131] In step S140, if the subsequence is a target feature in the target feature set, the computer system will use the enhanced template as the enhanced target knowledge template.

[0132] Taking the previous example, the subsequence is [relative value of peak height 1.2, relative value of trough depth 0.8]. Assume that the peak and trough morphological features represented by this subsequence are target features in the target feature set (used for epilepsy detection). Then, the previously obtained enhancement template [relative value of α-band energy 0.9, relative value of γ-band energy 1.1, relative value of peak height 1.2, relative value of trough depth 0.8] will be used as the enhanced target knowledge template.

[0133] This means that the template now possesses information related to the target features, and in subsequent operations (such as network tuning), it can be used as a sample with target features to participate in the optimization process of the monitoring data feature mining network. For example, in a neural network-based EEG modulation monitoring data feature mining model, the data in this enhanced target knowledge template can be used as positive examples input into the network, helping the network learn patterns related to the target features (features related to epileptic seizures), thereby improving the network's ability to recognize EEG modulation monitoring data with target features.

[0134] In step S150, if the subsequence is not a target feature in the target feature set, the computer system will use the enhanced template as an enhanced targetless knowledge template.

[0135] Suppose that in another scenario, the subsequence sampled from the first data item sequence is not a target feature in the target feature set. For example, if the subsequence is a feature related to EEG signal noise (and the noise feature is not part of the target feature set used for epilepsy detection), then the enhancement template containing this subsequence is still considered an enhancement template without a target.

[0136] Even though this enhanced template adds other features to the original targetless knowledge template, it is still used to represent EEG modulation monitoring data that does not have target features because the added subsequences are not target features. In the neural network model, the data in this enhanced targetless knowledge template can be used as a counterexample input to the network, helping the network learn patterns related to non-target features, thereby improving the network's ability to distinguish between EEG modulation monitoring data with and without target features.

[0137] In step S160, the computer system obtains a knowledge template library for the monitoring data feature mining network based on the original template library, the enhanced targeted knowledge templates, and the enhanced non-targeted knowledge templates.

[0138] The original template library contains targeted and non-targeted knowledge templates initially obtained from various sources (such as epilepsy patients and healthy individuals), providing basic feature representations of EEG modulation monitoring data. Enhanced targeted knowledge templates are obtained by adding subsequences related to target features to the original template library, enriching the content of knowledge templates with target features. Enhanced non-targeted knowledge templates are obtained by adding non-target feature-related subsequences to the original non-targeted knowledge templates, further refining the knowledge templates that do not possess target features.

[0139] The computer system integrates all templates from the original template library, enhanced targeted knowledge templates, and enhanced non-targeted knowledge templates to form a knowledge template library for the monitoring data feature mining network. This knowledge template library contains a rich variety of EEG modulation monitoring data feature representations, including both basic features obtained from the original data and features enhanced by adding subsequences.

[0140] In deep learning-based neural network models, template data in this knowledge template library serves as crucial input for network tuning and other operations. For example, when training a neural network for detecting epilepsy using EEG signals, target-oriented knowledge template data (including original and enhanced versions) in the knowledge template library can serve as positive examples, while targetless knowledge template data (including original and enhanced versions) can serve as negative examples. By continuously learning the feature patterns in these positive and negative examples, the network adjusts parameters such as weights, thereby improving its ability to mine target features (seizure-related features) from EEG modulation monitoring data. This knowledge template library provides a comprehensive foundation of feature information for the construction and optimization of the entire monitoring data feature mining network, ensuring that the network can accurately identify and mine target features in EEG modulation monitoring data.

[0141] As one implementation method, the maximum loading size of the monitoring data feature mining network is M, where M ≥ 1; based on this, after obtaining the knowledge template base of the monitoring data feature mining network, the method may further include:

[0142] Step S101: For the EEG modulation monitoring data of the knowledge template to be loaded into the monitoring data feature mining network, if the size of the EEG modulation monitoring data of the knowledge template is greater than the maximum loading size M, filter out the non-target data items in the EEG modulation monitoring data of the knowledge template to obtain the simplified data of the knowledge template.

[0143] Step S102: If the size of the simplified data of the knowledge template is not greater than the maximum loading size M, use the simplified data of the knowledge template as the loading data of the monitoring data feature mining network.

[0144] Step S103: If the size of the simplified data of the knowledge template is greater than the maximum loading size M, then extract EEG modulation monitoring data with a size not greater than the maximum loading size M from the simplified data of the knowledge template as the loading data of the monitoring data feature mining network.

[0145] In step S101, the computer system targets the EEG modulation monitoring data of the knowledge template to be loaded into the monitoring data feature mining network. If the size of the EEG modulation monitoring data of the knowledge template is greater than the maximum loading size M, the non-target data items in the EEG modulation monitoring data of the knowledge template are filtered out to obtain simplified data of the knowledge template.

[0146] In the deep learning-based EEG modulation monitoring data feature mining task, assume the maximum loading size M of the monitoring data feature mining network is 100 (here, 100 is just for illustrative purposes to represent a data volume limit). Taking EEG signal analysis as an example, the EEG modulation monitoring data in the knowledge template may contain data with multiple features, which can be represented by vectors or arrays.

[0147] For example, the EEG modulation monitoring data of a knowledge template contains a long feature vector, which is represented as [α band energy value 1, α band energy value 2, α band energy value 3, β band energy value 1, β band energy value 2, signal peak height 1, signal peak height 2, signal trough depth 1, signal trough depth 2, signal synchronicity index of different brain regions 1, signal synchronicity index of different brain regions 2, ...], and the total number of its data items is greater than 100.

[0148] Suppose the target features are alpha band energy characteristics and signal peak height characteristics relevant to epilepsy detection. The computer system needs to determine which data items are non-target data items. In this example, beta band energy values, signal trough depths, and indicators of signal synchronicity in different brain regions might be considered non-target data items.

[0149] The computer system identifies these non-target data items using specific algorithms or rules. For example, it can be determined based on the definition of target features when the knowledge template base was previously built, as well as the correlation analysis between the features and the target features. Then, the system filters out these non-target data items, obtaining simplified data, which may be simplified to [α-band energy value 1, α-band energy value 2, α-band energy value 3, signal peak height 1, signal peak height 2].

[0150] The purpose of this filtering operation is to remove non-target data items that may not be important for target feature mining when the amount of EEG modulation monitoring data in the knowledge template is too large, thereby reducing the data size so that it can adapt to the maximum loading size limit of the monitoring data feature mining network, while retaining as much key information related to the target features as possible so that the data can still be effectively used in subsequent operations (such as network tuning).

[0151] In step S102, if the size of the simplified data of the knowledge template is not greater than the maximum loading size M, the simplified data of the knowledge template is used as the loading data of the monitoring data feature mining network.

[0152] Continuing with the example above, if the simplified data obtained after filtering in step S101, such as [α band energy value 1, α band energy value 2, α band energy value 3, signal peak height 1, signal peak height 2], has a total number of data items less than or equal to 100 (i.e., the maximum loading size M).

[0153] In the context of using deep learning-based neural network models for feature mining of EEG modulation monitoring data, this simplified data can be directly used as the loading data for the monitoring data feature mining network. For example, in a multilayer perceptron (MLP) neural network, this simplified data can serve as input data for the input layer. The number of neurons in the input layer can be set according to the dimensionality of the simplified data, with each neuron receiving the value of one data item. Then, the network can process this input data according to pre-defined weights and activation functions, progressively performing feature mining and classification in the hidden and output layers (for classification tasks, such as distinguishing EEG signals in seizure and non-seizure states). This approach ensures that, while meeting network loading size constraints, data relevant to the target features can be effectively input into the monitoring data feature mining network, providing suitable data input for the normal operation of the network and accurate feature mining.

[0154] In step S103, if the size of the simplified data of the knowledge template is greater than the maximum loading size M, then the computer system extracts EEG modulation monitoring data with a size not greater than the maximum loading size M from the simplified data of the knowledge template as the loading data for the monitoring data feature mining network.

[0155] Suppose that in another scenario, the simplified data after filtering in step S101 is still large, for example, the simplified data is [α-band energy value 1, α-band energy value 2, α-band energy value 3, α-band energy value 4, α-band energy value 5, signal peak height 1, signal peak height 2, signal peak height 3, signal peak height 4, signal peak height 5, ...], and the total number of data items is still greater than 100 (maximum loading size M). The computer system needs to further extract some data as loading data. There are several ways to extract the data, such as in a certain order (e.g., from front to back) or based on the importance of the data (if the importance of the data items can be assessed in advance).

[0156] Assuming the data is extracted sequentially from front to back, yielding [α-band energy value 1, α-band energy value 2, α-band energy value 3, α-band energy value 4, α-band energy value 5, signal peak height 1, signal peak height 2, signal peak height 3, signal peak height 4], and the size of this extracted data is no greater than 100, then this data can be used as loading data for a monitoring data feature mining network. In the neural network model, this extracted data serves as input to the input layer, and the network performs feature mining operations based on this input data. Although this is an incomplete use of data (due to data size limitations), through reasonable extraction methods, it can still provide the network with information related to the target features to a certain extent, enabling the network to perform preliminary feature mining. While there may be certain limitations due to data incompleteness, in practical applications, this is an effective strategy to cope with data size constraints.

[0157] Through steps S101 to S103, the computer system can reasonably process the EEG modulation monitoring data in the knowledge template according to the maximum loading size M of the monitoring data feature mining network, ensuring that appropriate data can be loaded into the network, thereby providing effective data support for subsequent network tuning, feature mining and other operations. It also adapts to the hardware resource limitations of the network (such as memory limitations, and avoids memory overflow and other problems by limiting the loading data size).

[0158] As one implementation method, step S101, filtering out non-target data items in the EEG modulation monitoring data of the knowledge template to obtain simplified data of the knowledge template, may include:

[0159] Step S1011: Walk through each data item in the EEG modulation monitoring data of the knowledge template to obtain the identifier of the data item in the data dictionary; wherein, different data items in the data dictionary correspond to different identifiers;

[0160] Step S1012: Filter out data items in the EEG modulation monitoring data of the knowledge template whose identifiers do not belong to the target identifier range to obtain simplified data of the knowledge template; wherein, the target identifier range is the range determined based on the identifiers of each data item of the target feature.

[0161] In step S101, when the size of the EEG modulation monitoring data of the knowledge template is greater than the maximum loading size M, the computer system needs to filter out the non-target data items in the EEG modulation monitoring data of the knowledge template to obtain simplified data of the knowledge template. This process is implemented through sub-steps S1011 to S1012.

[0162] In step S1011, the computer system navigates through each data item in the EEG modulation monitoring data of the knowledge template to obtain the identifier of each data item in the data dictionary. Different data items in the data dictionary correspond to different identifiers.

[0163] In the context of deep learning-based EEG modulation monitoring data processing, the EEG modulation monitoring data of a knowledge template is a dataset containing multiple types of information. For example, taking EEG signal analysis for the diagnosis of a certain brain disease as an example, the EEG modulation monitoring data in the knowledge template may include the energy values ​​of EEG signals in different frequency bands (such as alpha, beta, and gamma bands), the peak and trough characteristics of EEG signals (such as peak height, trough depth, and peak spacing), and the signal synchronization characteristics between different brain regions.

[0164] Assuming the EEG modulation monitoring data is represented in array form, a simple example might be [alpha band energy value 0.3, beta band energy value 0.5, peak height 1.2, trough depth 0.8, frontal-temporal lobe signal synchronicity index 0.6].

[0165] A data dictionary is a reference for defining and encoding all possible data items. In this data dictionary, each data item has a unique identifier. For example, the identifier for alpha band energy value might be "001", the identifier for beta band energy value might be "002", the identifier for peak height might be "010", the identifier for trough depth might be "011", and the identifier for frontal-temporal lobe signal synchronicity index might be "100".

[0166] The computer system iterates through (i.e., walks through) each data item in the EEG modulation monitoring data within the knowledge template. Taking the array mentioned earlier as an example, the system first processes the alpha band energy value of 0.3, obtaining its corresponding identifier "001" by looking up the data dictionary; then it processes the beta band energy value of 0.5, obtaining its identifier "002"; then it processes the peak height of 1.2, obtaining the identifier "010"; and so on, until all data items have been processed. The purpose of this walk is to establish a connection between each data item and the data dictionary, thereby enabling the classification and manipulation of data items through identifiers. In subsequent steps, these identifiers will become important criteria for determining whether a data item is a target data item. From the perspective of machine learning models, when neural networks are used for feature mining of EEG modulation monitoring data, the data dictionary and identifiers help to standardize and classify the input data, allowing the network to better understand and process this data. For example, in the input layer, different data items can be assigned to different neurons or groups of neurons for processing based on identifiers, improving the network's efficiency and accuracy in processing data.

[0167] In step S1012, the computer system filters out data items in the EEG modulation monitoring data of the knowledge template whose identifiers do not belong to the target identifier range, thus obtaining simplified data of the knowledge template. The target identifier range is determined based on the identifiers of each data item possessed by the target feature.

[0168] Continuing with the example above, let's assume the target features are two data items related to the diagnosis of brain diseases: alpha band energy value and peak height. Based on the previous steps (such as the process of determining the target identifier range), the identifier corresponding to the alpha band energy value is "001," and its target identifier range is assumed to be "001~005" (possibly taking into account minor fluctuations in alpha band energy values ​​between different individuals or measurement errors); the identifier corresponding to the peak height is "010," and its target identifier range is assumed to be "010~015."

[0169] The computer system checks the identifier of each previously acquired data item. For the β-band energy value, the identifier is "002," which does not belong to the target identifier range of "001~005" for the α-band energy value, nor to the target identifier range of "010~015" for the peak height. Therefore, this data item will be filtered out. Similarly, the trough depth, with its identifier "011," is also not within the target identifier range and will be filtered out. The frontal-temporal lobe signal synchronicity index, with its identifier "100," does not belong to any target identifier range and will also be filtered out. After filtering, the remaining data items—α-band energy value and peak height—form the simplified data for the knowledge template, which is simplified to [α-band energy value 0.3, peak height 1.2].

[0170] The significance of this filtering operation lies in removing data items irrelevant to the target features when the size of EEG modulation monitoring data in the knowledge template becomes too large, thereby reducing the data size. In practical applications, this helps to process data related to the target features more effectively with limited resources (such as network maximum loading size limitations and computational resource limitations). For example, when using a deep learning-based neural network model for feature mining of EEG modulation monitoring data, an excessively large input data volume may lead to problems such as excessively long network training time and excessive memory consumption. By retaining only data items related to the target features as input through this filtering operation, the training efficiency of the network can be improved, while reducing unnecessary data interference, allowing the network to focus more on mining information related to the target features and improving the accuracy of tasks such as diagnosing brain diseases.

[0171] As one implementation, the method further includes a process for determining the range of target identifiers, including:

[0172] Step S101A: Obtain the identifiers of each data item of the target feature in the data dictionary;

[0173] Step S101B: Based on the identifiers and extended ranges corresponding to each data item of the target feature, determine the identifier range corresponding to each data item of the target feature. The extended range represents the numerical range of the identifier.

[0174] Step S101C: Determine the target identifier range based on the identifier range corresponding to each data item of the target feature.

[0175] In step S101A, the computer system obtains the identifiers of each data item of the target feature in the data dictionary. In this embodiment of the invention, the data dictionary is a reference for encoding and identifying various EEG modulation monitoring data features. Assuming we take EEG signal analysis for detecting specific brain diseases as an example, EEG signals have multiple features, such as energy values ​​in different frequency bands (alpha, beta, gamma, etc.), peak and trough characteristics of the signal (peak height, trough depth, peak spacing, etc.), and signal synchronization characteristics between different brain regions. For example, the data dictionary may be in tabular form, where each row corresponds to a data item, and each column contains different attributes of the data item, such as name, identifier, data type, and value range. For the data item "alpha band energy value," there may be a corresponding identifier in the data dictionary, for example, "001"; the identifier corresponding to the beta band energy value is "002"; the identifier corresponding to the peak height is "010," and so on.

[0176] The computer system iterates through each data item in the target feature, then looks up and retrieves their corresponding identifiers from the data dictionary. For example, if the target features are the alpha band energy value and peak height associated with a specific brain disease, the computer system would retrieve their identifiers "001" and "010" from the data dictionary. The purpose of this step is to provide the foundational elements for subsequently determining the range of target identifiers. By clearly defining the identifiers for the target feature data items, it is possible to accurately locate and distinguish data related to these target features within the entire data system.

[0177] In step S101B, the computer system determines the identifier range corresponding to each data item of the target feature based on the identifier and extension range corresponding to each data item of the target feature. The extension range represents the numerical range of the identifier.

[0178] Continuing with the example above, the identifier for the alpha band energy value in the data dictionary is "001". Let's assume its range is determined based on statistical analysis of a large amount of EEG monitoring data or specific medical knowledge. For instance, in normal EEG signals, the fluctuation range of the alpha band energy value is relatively stable, and the range of the corresponding identifier "001" might be defined as "001~005". This range represents the possible variation range of the identifier related to the alpha band energy value, which may be due to minor differences between individuals or measurement errors.

[0179] For the data item of peak height, its identifier is "010". Assuming that based on the variation pattern of peak height under different EEG states and factors such as measurement accuracy, its extended range is determined to be "010~015".

[0180] This method of determining identifier ranges is based on a deep understanding of the target feature data items and the statistical analysis of related data. For example, when using machine learning-based neural network models for feature mining of EEG modulation monitoring data, these identifier ranges can help the network more accurately identify and process data related to the target features. If the network receives a data item with the identifier "003", it can determine that this data item may be related to the alpha band energy value based on the previously determined identifier range ("001~005"), and thus perform appropriate processing.

[0181] In step S101C, the computer system determines the target identifier range based on the identifier range corresponding to each data item of the target feature.

[0182] In the preceding steps, we have determined the identifier ranges corresponding to each target feature data item. For example, the identifier range for the α-band energy value is "001~005", and the identifier range for the peak height is "010~015". The computer system integrates these identifier ranges to determine the target identifier range.

[0183] In this example, the target identifier range could be the union of "001~005" and "010~015", i.e., "001~005, 010~015". This target identifier range covers the identifier range of all data items related to the target feature in the data dictionary.

[0184] This target identifier range is crucial for subsequent operations. For example, when filtering out non-target data items in the EEG modulation monitoring data of the knowledge template in step S101, the computer system can determine whether to retain a data item by checking whether its identifier in the data dictionary falls within the target identifier range. If a data item's identifier is outside this target identifier range, such as "020", then this data item may be judged as a non-target data item and filtered out. This ensures that only data related to the target features is retained when processing EEG modulation monitoring data, improving the efficiency and accuracy of data processing. It provides cleaner and more targeted data input for the deep learning-based EEG modulation monitoring data feature mining network, helping to improve the network's ability to mine target features and distinguish between different EEG states (such as normal and disease states).

[0185] As one implementation method, in step S400, after adjusting the pruned monitoring data feature mining network through the knowledge template library to obtain a converged monitoring data feature mining network, the method may further include: reducing the floating-point value of one or more weight parameters in the converged monitoring data feature mining network to obtain a converged monitoring data feature mining network with reduced floating-point value.

[0186] When the computer system executes the deep learning-based EEG modulation monitoring data feature mining method, in step S400, the pruned monitoring data feature mining network is adjusted through the knowledge template library to obtain a converged monitoring data feature mining network. After that, the floating point of one or more weight parameters in the converged monitoring data feature mining network is reduced to obtain a converged monitoring data feature mining network with reduced floating point.

[0187] In the context of using deep learning-based neural network models for feature mining of EEG modulation monitoring data, the monitoring data feature mining network is a complex structure. The weight parameters within it determine the connection strength between neurons and the flow and transformation of data within the network. After the network is tuned and reaches convergence, its weight parameters have specific values, typically represented as floating-point numbers.

[0188] For example, suppose the feature mining network for monitoring data in the convergent state is a multilayer perceptron (MLP), where the connection weights between neurons in one hidden layer and neurons in the next layer are floating-point numbers such as [0.123456, -0.234567, 0.345678]. The computer system performs a de-floating-point operation on these weights.

[0189] This floating-point reduction operation has multiple implications and benefits. First, from the perspective of data storage and computational resources, reducing the floating-point number of weight parameters can decrease the storage space requirements. For example, reducing the floating-point number from double precision (64 bits) to single precision (32 bits) can significantly reduce storage space requirements. Take a simple weight parameter of 0.123456789 as an example: if stored in double precision, it requires 64 bits of storage space; however, if reduced to single precision, only 32 bits are needed. In large-scale neural networks, especially those processing complex data such as EEG modulation monitoring, there are numerous weight parameters. By reducing the number of floating-point bits, the overall storage requirements of the network can be significantly reduced. This is crucial for resource-constrained devices (such as some portable EEG monitoring devices), enabling the storage and operation of network models within limited storage space.

[0190] Secondly, from a computational efficiency perspective, lower-precision floating-point calculations are generally faster than high-precision floating-point calculations. During the forward and backward propagation processes of a neural network, a large number of multiplication and addition operations are required on the weight parameters. Taking a simple neural network layer as an example, assuming the input vector is [x1, x2, x3] and the weight matrix is... ,in The weight parameter is used to calculate the output vector y. If the floating-point values ​​of the weight parameters are reduced, for example, from double precision to single precision, the computational units (such as CPUs or GPUs) can complete the calculations faster when performing these multiplication and addition operations, thereby improving the overall network speed.

[0191] However, reducing the floating-point precision of weight parameters comes at a cost. The reduced numerical precision may introduce errors. For example, in feature mining of EEG signals, excessively reducing the precision of weight parameters could decrease the network's ability to recognize subtle changes in EEG signal features. Suppose a weak, disease-related feature change in the EEG signal can be accurately identified in a network with high-precision weight parameters, but in a network with reduced floating-point precision, this weak feature might be overlooked or misjudged due to the loss of weight parameter precision.

[0192] To balance this trade-off between precision loss and resource conservation, computer systems need to determine floating-point reduction strategies based on specific application scenarios and requirements. For example, if the primary purpose of EEG modulation monitoring is to quickly classify obvious EEG signal features (such as distinguishing between normal EEG signals and those during epileptic seizures, where the changes in EEG signal features are more pronounced), then a significant reduction in floating-point numbers can be achieved in exchange for faster computation speed and lower storage requirements. However, if the goal is to precisely analyze subtle EEG signal features (such as studying minute differences in EEG signals under different cognitive states), then a more cautious approach to reducing floating-point numbers may be necessary, or specialized quantization techniques may be employed to minimize precision loss.

[0193] In practice, computer systems can employ various methods to reduce the floating-point representation of weight parameters. A common approach is quantization. For example, floating-point numbers can be mapped to a smaller numerical range according to certain rules and represented using integers. Suppose the weight parameter's value range is [-1, 1]. This range can be divided into several intervals, such as [-1, -0.5), [-0.5, 0), [0, 0.5), and [0.5, 1]. Then, the floating-point numbers within these intervals can be represented by integers ~3, ~2, ~1, 0, 1, 2, 3, etc. In this way, the original floating-point weight parameter is converted to an integer representation, significantly reducing data storage and computational complexity.

[0194] In summary, reducing the floating-point values ​​of weight parameters in the feature mining network of converged monitoring data is a means of balancing resource utilization and model performance. Through a reasonable floating-point reduction strategy, the storage and computational efficiency of the network can be improved while meeting the requirements of EEG modulation monitoring tasks, providing a more optimized solution for practical EEG modulation monitoring applications.

[0195] As one implementation method, step S500, determining whether the target feature exists in the brainwave modulation monitoring data to be mined based on the data feature vector of the data feature vector of the brainwave modulation monitoring data to be mined, may include:

[0196] Step S510: Obtain the target feature set; the target feature set includes multiple target data item sequences and a data item attribute sequence corresponding to each target data item sequence;

[0197] Step S520: Find the matching target data item sequence that matches the data feature vector in multiple target data item sequences, and extract the target data item attribute sequence corresponding to the matching target data item sequence;

[0198] Step S530: Identify the feature categories that match the data feature vector based on the matching target data item sequence and the target data item attribute sequence;

[0199] Step S540: Obtain the monitoring data attributes corresponding to the feature categories as the monitoring data attributes for mining;

[0200] Step S550: If the target data item sequence contains mining monitoring data attributes, then it is determined that the EEG modulation monitoring data to be mined has target features, and the matching target data item sequence containing mining monitoring data attributes is taken as the target mining monitoring data item.

[0201] In step S500, the computer system determines whether the target feature exists in the brainwave modulation monitoring data to be mined based on the data feature vector of the data to be mined. This process is implemented through sub-steps S510 to S550.

[0202] In step S510, the computer system acquires a target feature set. The target feature set includes multiple sequences of target data items and a sequence of data item attributes corresponding to each sequence of target data items.

[0203] In feature mining of EEG modulation monitoring data, taking the detection of features in EEG signals related to epileptic seizures as an example, the target feature set may contain multiple EEG signal feature patterns associated with epileptic seizures.

[0204] Suppose that one of the target data item sequences is a sequence concerning the changes in frequency band energy of EEG signals. This sequence can be represented as [relative value of alpha band energy before seizure, relative value of alpha band energy during seizure, relative value of alpha band energy after seizure], where the relative value is relative to the energy value under normal conditions. For example, the relative value of alpha band energy under normal conditions is 1, it may become 0.9 before a seizure, decrease to 0.5 during a seizure, and recover to 0.8 after a seizure. The data item attribute sequence corresponding to this target data item sequence may be [threshold range of relative energy value, threshold range of relative energy value, threshold range of relative energy value], for example, [pre-seizure: 0.8~1.1, during seizure: 0~0.6, after seizure: 0.7~1]. These threshold ranges are obtained based on statistical analysis of EEG signals from a large number of epilepsy patients and are used to define whether the relative values ​​of alpha band energy at different stages conform to the characteristic patterns associated with epileptic seizures. Let's assume another target data item sequence concerns the signal synchronicity between different brain regions, such as [pre-ictal signal synchronicity index between the frontal and temporal lobes, intra-ictal signal synchronicity index between the frontal and temporal lobes, and post-ictal signal synchronicity index between the frontal and temporal lobes]. The data item attribute sequence might be [threshold range of synchronicity index, threshold range of synchronicity index, threshold range of synchronicity index], for example, [pre-ictal: 0.6~0.9, intra-ictal: 0.2~0.5, post-ictal: 0.6~0.9]. Here, the signal synchronicity index can be a value obtained by calculating the correlation between EEG signals in the frontal and temporal lobes, used to measure the degree of synchronicity between the signals in the two brain regions. These target data item sequences and their corresponding attribute sequences together constitute the target feature set, providing a comprehensive reference standard for subsequently determining whether the EEG modulation monitoring data to be mined possesses the target features.

[0205] In step S520, the computer system searches for a matching target data item sequence that matches the data feature vector among multiple target data item sequences, and extracts the target data item attribute sequence corresponding to the matching target data item sequence. Assume that in the previous steps, the computer system obtains the data feature vector of the EEG modulation monitoring data to be mined through the representation information extraction component of the convergent monitoring data feature mining network as [0.85 (relative value of α band energy before the attack), 0.4 (relative value of α band energy during the attack), 0.75 (relative value of α band energy after the attack), 0.7 (signal synchronicity index of frontal and temporal lobes before the attack), 0.3 (signal synchronicity index of frontal and temporal lobes during the attack), 0.7 (signal synchronicity index of frontal and temporal lobes after the attack)].

[0206] The computer system compares this data feature vector with the sequence of target data items in the target feature set. For the target data item sequence concerning the frequency band energy changes of EEG signals [relative value of alpha band energy before the attack, relative value of alpha band energy during the attack, relative value of alpha band energy after the attack], the corresponding value in the data feature vector is within its corresponding threshold range (pre-attack: 0.8~1.1, during the attack: 0~0.6, after the attack: 0.7~1).

[0207] For the target data item sequence regarding signal synchronicity between different brain regions [pre-ictal signal synchronicity index between frontal and temporal lobes, intraictal signal synchronicity index between frontal and temporal lobes, post-ictal signal synchronicity index between frontal and temporal lobes], the corresponding values ​​in the data feature vector are also within their corresponding threshold ranges (pre-ictal: 0.6~0.9, intraictal: 0.2~0.5, post-ictal: 0.6~0.9).

[0208] Therefore, both of these target data item sequences can be considered as matching target data item sequences that match the data feature vector. Then, the computer system extracts the target data item attribute sequences corresponding to these two matching target data item sequences. For the matching target data item sequence of frequency band energy change, the corresponding attribute sequence is [before attack: 0.8~1.1, during attack: 0~0.6, after attack: 0.7~1]; for the matching target data item sequence of signal synchronization, the corresponding attribute sequence is [before attack: 0.6~0.9, during attack: 0.2~0.5, after attack: 0.6~0.9].

[0209] In step S530, the computer system identifies the feature category that matches the data feature vector based on the matching target data item sequence and the target data item attribute sequence.

[0210] Continuing with the example above, the computer system has identified two matching target data item sequences and their corresponding attribute sequences. For the matching target data item sequence and its attribute sequence representing frequency band energy changes, it represents a category of EEG signal energy features associated with epileptic seizures. By comparing the values ​​in the data feature vector with the threshold ranges in the attribute sequence, the computer system can determine that this data feature vector conforms to a characteristic pattern associated with epileptic seizures in terms of frequency band energy changes.

[0211] For a sequence of target data items and their attribute sequences that match signal synchronization, it represents a characteristic category of signal synchronization between brain regions associated with epileptic seizures. Similarly, by comparing the values ​​in the data feature vector with the threshold ranges in the attribute sequence, it can be determined that this data feature vector also conforms to the characteristic pattern associated with epileptic seizures in terms of signal synchronization.

[0212] Therefore, based on these matching target data item sequences and target data item attribute sequences, the computer system identifies the feature categories of the data feature vector matching as epileptic seizure-related frequency band energy change features and brain region signal synchronization features. This identification helps to associate the information in the data feature vector with specific EEG signal feature patterns, providing a more detailed basis for subsequent judgment on whether the EEG modulation monitoring data to be mined has the target features.

[0213] In step S540, the computer system obtains the monitoring data attributes corresponding to the feature categories as the monitoring data attributes for mining.

[0214] For the frequency band energy change characteristics identified above as related to epileptic seizures, the corresponding monitoring data attribute may be an "epilepsy seizure risk assessment index." This index can be a value calculated based on the degree of frequency band energy change and the performance at different stages of the seizure. For example, if the relative value of the alpha band energy is lower during a seizure, and the energy change before and after the seizure is more significant, this "epilepsy seizure risk assessment index" may be higher.

[0215] For categories of brain region signal synchronicity characteristics associated with epileptic seizures, the corresponding monitoring data attribute might be a "brain functional connectivity stability index." This index reflects the impact of changes in signal synchronicity between the frontal and temporal lobes during an epileptic seizure on the overall stability of brain functional connectivity. For example, the lower the signal synchronicity during a seizure, the lower this "brain functional connectivity stability index" is likely to be, indicating that brain functional connectivity is significantly affected during an epileptic seizure.

[0216] These data mining and monitoring attributes are key attributes related to EEG modulation monitoring, which are further extracted from feature categories. They can provide a more direct basis for judging whether the EEG modulation monitoring data to be mined has the target features.

[0217] In step S550, if the computer system determines that the target data item sequence contains mining monitoring data attributes, then the target features are found in the EEG modulation monitoring data to be mined, and the matching target data item sequence containing mining monitoring data attributes is taken as the target mining monitoring data item.

[0218] Taking the "epilepsy seizure risk assessment index" as an example of the data mining and monitoring data attribute, since it is related to the frequency band energy change feature category associated with epileptic seizures, and the target data item sequence corresponding to this feature category contains content related to the relative values ​​of α-band energy at different stages of epileptic seizures, it can be determined that the EEG modulation monitoring data to be mined contains target features related to epileptic seizures. Furthermore, the matching target data item sequence containing the "epilepsy seizure risk assessment index" data mining and monitoring data attribute (i.e., the target data item sequence regarding frequency band energy changes in EEG signals) is taken as the target data mining and monitoring data item.

[0219] Similarly, for the "brain functional connectivity stability index" attribute in the monitoring data, since it is related to the brain region signal synchronization feature category associated with epileptic seizures, and the target data item sequence corresponding to this feature category includes content related to signal synchronization indices at different stages of frontal and temporal lobe seizures, it can also be determined that the EEG modulation monitoring data to be mined contains target features related to epileptic seizures. The matching target data item sequence containing the "brain functional connectivity stability index" attribute (i.e., the target data item sequence regarding signal synchronization between different brain regions) is also used as the target monitoring data item. Through this series of steps, the computer system can accurately determine whether the data to be mined for EEG modulation monitoring possesses the target feature based on the data feature vector. This judgment is of great significance for applications such as disease diagnosis and treatment effect evaluation in EEG modulation monitoring, and can provide a strong basis for related medical decisions.

[0220] Figure 2 A hardware entity diagram of a computer system provided as an embodiment of the present invention, such as... Figure 2 As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

Claims

1. A method for feature mining of EEG modulation monitoring data based on deep learning, characterized in that, The method includes: Obtaining the knowledge template library of the monitoring data feature mining network specifically includes: obtaining the original template library of the monitoring data feature mining network, wherein the original template library includes one or more original target-oriented knowledge templates and one or more original targetless knowledge templates; sampling one or more target features from the target feature set to be mined, combining the one or more target features to obtain a first data item sequence; sampling a sub-sequence from the first data item sequence, and adding the sub-sequence to the original targetless knowledge template to obtain an enhanced template; wherein the sub-sequence is one or more uninterrupted data items in the first data item sequence; if the If the subsequence is a target feature in the target feature set, the enhanced template is used as an enhanced target-based knowledge template; if the subsequence is not a target feature in the target feature set, the enhanced template is used as an enhanced targetless knowledge template; based on the original template library, the enhanced target-based knowledge template, and the enhanced targetless knowledge template, the knowledge template library of the monitoring data feature mining network is obtained; the knowledge template library includes one or more target-based knowledge templates and one or more targetless knowledge templates, the target-based knowledge templates represent EEG modulation monitoring data with target features, and the targetless knowledge templates represent EEG modulation monitoring data without the target features; The monitoring data feature mining network is calibrated using the knowledge template library to obtain an initially calibrated monitoring data feature mining network. The network configuration variables of the initially calibrated monitoring data feature mining network are pruned to obtain a pruned monitoring data feature mining network; wherein the memory usage of the pruned monitoring data feature mining network is less than that of the initially calibrated monitoring data feature mining network; specifically, this includes: for the first network component in the initially calibrated monitoring data feature mining network, performing eigenvalue splitting on the weight tensor of the first network component to obtain the product of a first rotation tensor, an eigenvalue tensor, and a second rotation tensor; Among the eigenvalues ​​of the eigenvalue tensor, determining the target eigenvalues ​​that meet the first requirement specifically includes: sorting the eigenvalues ​​of the eigenvalue tensor in descending order to obtain an eigenvalue vector; determining the first p eigenvalues ​​in the eigenvalue vector to obtain p target eigenvalues, where p ≥ 2; In the first rotation tensor, the eigenvalue tensor, and the second rotation tensor, configuration variables corresponding to the target eigenvalue are extracted from each to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor. The configuration variables of the first network component in the pruned monitoring data feature mining network are composed of the pruned first rotation tensor, the pruned feature value tensor, and the pruned second rotation tensor. The pruned monitoring data feature mining network is tuned using the knowledge template library to obtain a converged monitoring data feature mining network. The representation information extraction component of the monitoring data feature mining network based on the convergence state obtains the data feature vector of the EEG modulation monitoring data to be mined, and determines whether the target feature is present in the EEG modulation monitoring data to be mined based on the data feature vector of the EEG modulation monitoring data to be mined.

2. The method according to claim 1, characterized in that, The first network component's weight tensor includes i*j weight parameters, the first rotation tensor includes i*i weight parameters, the eigenvalue tensor includes i*j weight parameters, the second rotation tensor includes j*j weight parameters, and the number of target eigenvalues ​​is p, where i≥2; j≥2; p≥2. The step of extracting configuration variables corresponding to the target feature value from the first rotation tensor, the eigenvalue tensor, and the second rotation tensor to obtain the pruned first rotation tensor, the pruned eigenvalue tensor, and the pruned second rotation tensor includes: Based on the target feature value, p columns of weight parameters corresponding to the target feature value are obtained in the first rotation tensor to obtain the pruned first rotation tensor, which includes i*p weight parameters. Based on the target feature value, the pruned feature value tensor is obtained, and the pruned feature value tensor includes p*p weight parameters. Based on the target feature value, obtain the p rows of weight parameters corresponding to the target feature value in the second rotation tensor to obtain the pruned second rotation tensor. The pruned second rotation tensor includes p*j weight parameters. The configuration variable in each feature value tensor corresponds to a column of weight parameters in the first rotation tensor and a row of weight parameters in the second rotation tensor.

3. The method according to claim 1, characterized in that, The maximum loading size of the monitoring data feature mining network is M, where M≥1; After obtaining the knowledge template library of the monitoring data feature mining network, the method further includes: For the EEG modulation monitoring data of the knowledge template to be loaded into the monitoring data feature mining network, if the size of the EEG modulation monitoring data of the knowledge template is greater than the maximum loading size M, the non-target data items in the EEG modulation monitoring data of the knowledge template are filtered out to obtain the simplified data of the knowledge template. If the size of the simplified data of the knowledge template is not greater than the maximum loading size M, the simplified data of the knowledge template will be used as the loading data of the monitoring data feature mining network. If the size of the simplified data of the knowledge template is greater than the maximum loading size M, then EEG modulation monitoring data with a size not greater than the maximum loading size M are extracted from the simplified data of the knowledge template as the loading data of the monitoring data feature mining network.

4. The method according to claim 3, characterized in that, The process of filtering out non-target data items from the EEG modulation monitoring data of the knowledge template to obtain simplified data of the knowledge template includes: The data items in the EEG modulation monitoring data of the knowledge template are walked through to obtain the identifiers of the data items in the data dictionary; wherein, different data items in the data dictionary correspond to different identifiers; In the EEG modulation monitoring data of the knowledge template, data items whose identifiers do not belong to the target identifier range are filtered out to obtain simplified data of the knowledge template; wherein, the target identifier range is the range determined based on the identifiers of each data item of the target feature.

5. The method according to claim 4, characterized in that, The method further includes: Obtain the identifiers of each data item of the target feature in the data dictionary; Based on the identifiers and extended ranges corresponding to each data item of the target feature, the identifier range corresponding to each data item of the target feature is determined, wherein the extended range represents the numerical range of the identifier; The target identifier range is determined based on the identifier range corresponding to each data item of the target feature.

6. The method according to claim 1, characterized in that, After adjusting the pruned monitoring data feature mining network using the knowledge template base to obtain a converged monitoring data feature mining network, the process further includes: The floating-point values ​​of one or more weight parameters in the converged monitoring data feature mining network are reduced to obtain the converged monitoring data feature mining network with reduced floating-point values.

7. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 6.

Images